Applicability Assessment of the 1998–2018 CLDAS Multi-Source Precipitation Fusion Dataset over China

As an important variable of the land surface, precipitation directly affects mass and energy exchanges between the atmosphere, biosphere, and lithosphere. Precipitation is a crucial variable in the simulation of soil moisture and runoff in modeling studies of land surface processes (Chahine, 1992; Shi et al., 2011; Liu et al., 2019). Precipitation is therefore a key research topic in climate change, agricultural development, and water use and management (Ren et al., 2015; Li et al., 2020; Miao and Wang, 2020).

Precipitation data can be obtained from station observations, remote sensing retrievals, and numerical simulations. These methods all have strengths and weaknesses and therefore one way to obtain a high-quality precipitation dataset is to merge data from various sources. There are currently a number of fusion datasets for precipitation. Since the 1990s, the technology of satellite data fusion has led to the development of the NASA Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis dataset (Huffman et al., 2007; Wang et al., 2017), the NOAA Climate Prediction Center (CPC) morphing technique (CMORPH) multisatellite precipitation dataset (Joyce et al., 2004), the Japan Aerospace Exploration Agency Global Satellite Mapping of Precipitation dataset (Ushio et al., 2009), and the Global Precipitation Measurement (GPM) precipitation fusion dataset (Hou et al., 2014). These satellite precipitation datasets provide support for global monitoring of heavy precipitation and extreme weather events. However, large systematic and random biases still exist in these satellite retrieval products, which need to be corrected based on in situ measurements. The accuracy of precipitation products can be further improved after data correction and fusion. The NOAA/CPC Unified (CPCU) global gridded precipitation dataset (Chen et al., 2008) is a good example of this type of fusion data. However, traditional hourly rain gauges and automatic stations do not accurately measure solid precipitation, except for a few stations with weighing instruments. The ability to retrieve solid precipitation from satellite remote sensing is still not satisfactory.

Progress has been made in the development of precipitation fusion datasets in China. Using the inverse distance ratio weighting method, a precipitation fusion dataset has been produced by the National Meteorological Center based on numerical precipitation forecasts and observations at > 30,000 regional automatic meteorological stations. This dataset has a spatial resolution of 5 km and a temporal resolution of 1 h. The East Asian Multi-Satellite Integrated Precipitation dataset is produced by the National Meteorological Information Center (NMIC) based on observations from multiple satellites [e.g., Fengyun-3B (FY-3B), NOAA-18/19, TRMM, and Meteorological Operational (MetOp)-A; Xu et al., 2015]. The data fusion technique probability density function (PDF) + the optimum interpolation (OI) is applied to produce hourly, 10-km resolution fusion data based on both surface observations and satellite retrievals (Pan et al., 2012). Although this dataset has been widely used, severe problems have been found for solid precipitation during the winter. The PDF + Bayesian model averaging (BMA) + OI fusion technique was developed to obtain hourly precipitation at 1-km/5-km resolution based on surface observations, satellite retrievals, and radar measurements (Pan et al., 2015). The quality of this new dataset is improved, but the problem related to solid precipitation in the winter still exists. The numerical model has some ability to simulate solid precipitation (Trenberth and Olson, 1988). Precipitation datasets such as the Modern-Era Retrospective analysis for Research and Applications (MERRA), MERRA2, ECMWF Interim reanalysis (ERA-Interim), Climate Forecast System Reanalysis, and the Multi-Source Weighted-Ensemble Precipitation currently contain both numerical model precipitation and in situ precipitation.

This paper proposes a method to obtain hourly precipitation data based on in situ observations on various temporal scales, satellite retrievals of precipitation, and reanalysis products. The final hourly precipitation product effectively makes up for the lack of hourly in situ observations over the icy regions of China. The Space–Time Multiscale Variational Analysis System (STMAS) method is implemented to merge the in situ observations and background data to produce a long-term, gridded, hourly precipitation fusion dataset (CLDAS-Prcp) that is stable over all four seasons.

The CLDAS was developed in four stages from CLDAS-V1.0 to CLDAS-V4.0 by a team led by Chunxiang Shi at the NMIC. CLDAS-V2.0 has been released and applied in operational numerical weather forecasts, agriculture, ecological hydrology, and drought monitoring (Shi et al., 2011; Han et al., 2019). The atmospheric driving data of CLDAS-V2.0 include the temperature, pressure, wind speed, precipitation, humidity, and solar shortwave radiation. The land surface models include the Community Land Model (CLM; Oleson et al., 2004), the Common Land Model (CoLM; Dai et al., 2003), and the community Noah land surface model with multiparameterization options (Noah-MP; Niu et al., 2011). The soil moisture content, soil temperature, and surface skin temperature simulated by CLDAS have been widely used by meteorological bureaus, colleges, and research institutes (Han et al., 2017; Sun et al., 2017). The CLDAS-Prcp precipitation data are produced from CLDAS.

The CMA in situ observations of precipitation can be divided into two periods by the observational instruments used. Before the year 2000, precipitation was measured by cylindrical rain gauges, tipping bucket rain gauges, or siphonic rain gauges, whereas since the 2000s, automatic weather stations have gradually replaced the older rain gauges. Hourly in situ observations are now largely obtained automatically. However, hourly observations are not available over icy areas in winter from either rain gauges or automatic meteorological stations and the only data available are the daily precipitation observational data measured by the station duty staff at the local meteorological administration. The CMA has recently begun to deploy weighing-type precipitation sensors to measure solid precipitation, but the number of such stations is still small.

Hourly precipitation data from 2400 national meteorological stations from 1998 to 2018 were extracted from the China National Surface Weather Station Hourly Precipitation Data (V2.0) dataset provided by the NMIC. These data are mainly derived from meteorological observation reports, digital self-recorded paper images, and automatic station data. Strict quality control has been conducted on this dataset (Ren et al., 2007).

The hourly precipitation data of 30,000–60,000 regional meteorological stations from 2008 to 2018 were extracted from the China Integrated Meteorological Information Service System (CIMISS). These data are marked with a quality control code that has passed strict quality control measures (Zhao et al., 2017).

The in situ observations were extracted from the China National Surface Weather Station Daily Precipitation Data (V3.0) dataset provided by the NMIC. Daily precipitation during the time periods 2000–0800, 0800–2000, and 2000–2000 Beijing Time (BT) were transformed to UTC time to match the numerical outputs and satellite retrievals (Li et al., 2015). These daily precipitation data contain information about solid precipitation. The CMORPH remote sensing retrievals and MERRA2 reanalysis products were also used. Table 1 lists the major features of these datasets.

The STMAS method is derived from the local analysis and prediction system developed at the Earth System Research Laboratory of NOAA. The local analysis and prediction system is used for the analysis and fusion of data from multiple sources, in which the STMAS method has been widely applied for the analysis of temperature, clouds, and the earth’s surface (e.g., dew temperature, pressure, and winds) (Xie et al., 2011). The STMAS method introduces a multiple grid approach to data assimilation (Zhang et al., 2014). The study area is divided into various sub-regions with grids from coarse to fine resolutions (Fig. 1) and different grids are analyzed to merge the observations with the background information. This method can make full use of observations and can also improve the analysis of heterogeneous observations (Zhang et al., 2014).

Figure 1.  Schematic diagram of the STMAS method (3/4DVARs: three- or four-dimensional variational data assimilation; Zhang et al., 2014).

In contrast with three-dimensional variational analysis, a wavelength is specified in STMAS and this is used to determine the correlation scale between grids. In this way, the STMAS has an internal correlation and thus does not need covariance to obtain analytical solutions. Therefore, the background bias covariance matrix (B matrix) can be simplified to a diagonal matrix. In STMAS, the target function on each grid is expressed as:

where Y is the deviation of the observation from the background (i.e., ${{Y}}={{Y}}^{\mathrm{o}\mathrm{b}\mathrm{s}}-{{H}}{{X}}^{\rm{b}})$, ${{Y}}^{\mathrm{o}\mathrm{b}\mathrm{s}}$ is the observed vector, ${{X}}^{\rm{b}}$ is the background vector, and H is the observation operator (i.e., the bilinear interpolation operator from the background field to the observation point); n indicates the nth grid and N is the total number of grids. The X represents the corrected vector relative to the model vector and O is the covariance matrix of the measurement biases.

where ${{X}}^{(n-1)}$ represents the results of analysis on the up-scale grid; ${{J}}^{(n)}$ on various grids is obtained first and ${{X}}^{\left(n\right)}$ is obtained through the minimization of ${{J}}^{(n)}$. Increment fields on multiscale grids are then obtained and the final analysis result is the sum of the background and the increment, that is, ${{X}}^{\rm{a}}={{X}}^{\rm{b}}+\sum _{{n}=1}^{N}{{X}}^{(n)}$.

Figure 2 shows the flowchart for the production of CLDAS-Prcp. Considering the low capability for the retrieval of solid precipitation in CMORPH, especially in the winter (Pan et al., 2012), MERRA2 was selected for statistical downscaling of the daily precipitation. According to Jordan (1991), 2.5°C was specified as the threshold to distinguish between rain and snow. If the daily minimum temperature at a specific station < 2.5°C, then the downscaled hourly data were used; otherwise, the real observations were used. Using this method, a new hourly precipitation dataset was obtained to make up for the missing data in icy areas.

Figure 2.  Flowchart for the production of the CLDAS-Prcp dataset.

The selection of the background field is crucial in the success of data fusion. The CMORPH and MERRA2 data were remapped to grids at a resolution of 0.0625° using the nearest neighbor interpolation method. The precipitation at a specific grid was set to the value from the MERRA2 dataset if the daily minimum temperature at this grid < 2.5°C; otherwise, the precipitation in this grid was set to the value from the CMORPH dataset. In this way, we were able to take advantage of both the satellite retrievals and the reanalysis products to obtain the background field. The gridded temperature data were extracted from CLDAS-V2.0 (Shi et al., 2014).

The precipitation dataset was evaluated qualitatively and quantitatively. Qualitative evaluation was conducted to examine whether these data can reasonably reflect the spatial pattern of precipitation. Quantitatively, this dataset was verified against in situ observations based on the bias, the root-mean-square error (RMSE), and the correlation coefficient.

These data were also used as forcing data for the Noah-MP land surface model to indirectly verify their quality. Noah-MP was developed based on the Noah land surface model and provides multiple parameterization options for biogeophysical processes, such as dynamic vegetation, runoff, and groundwater (Yang et al., 2011). Users can configure the parameterization scheme according to the simulation need. We chose the default scheme for soil moisture and snow depth simulation.

The CLDAS-Prcp data were evaluated using in situ observational data from the Ministry of Water Resources (MWR) and the CMA, which compared them with the precipitation from the Global Land Data Assimilation System version 2 (GLDAS-V2.1) and the GPM. The simulation result driven by CLDAS-Prcp was evaluated by the in situ soil moisture content and snow depth.

The CLDAS-V2.0 precipitation data were blended with the CMORPH and CMA hourly precipitation data, and cover East Asia from 2008 to 2018. They can be downloaded from the NMIC (http://data.cma.cn/; Pan et al., 2012). The GLDAS-V2.1 precipitation data were disaggregated by the NOAA/NCEP Global Data Assimilation System atmospheric analysis fields and the Global Precipitation Climatology Project precipitation fields. The GLDAS-V2.1 dataset covers the period from the year 2000 to the present day with a resolution of 3 h and 0.25°. It can be downloaded from the NASA website (https://hydro1.gesdisc.eosdis.nasa.gov/data/GLDAS/; Rodell et al., 2004). The GPM is a new generation of global satellite precipitation observation programs jointly proposed by NASA and the Japan Aerospace Exploration Agency. The program is based on the integrated multisatellite retrievals for the GPM, which can obtain precipitation data with a time resolution of 30 min and can also be downloaded from the NASA website (https://gpm1.gesdisc.eosdis.nasa.gov/data/GPM_L3/; Anjum et al., 2018).

Because the CMA observational precipitation data have been merged in CLDAS-Prcp, 2380 national meteorological stations were selected for dependent tests and compared with the GLDAS-V2.1, MERRA2, and CMORPH datasets. The MWR precipitation observational data were selected for independent tests and compared with the GPM and CLDAS-V2.0 precipitation. The daily MWR precipitation data were downloaded from the CIMISS website; missing and abnormal values were eliminated for quality control purposes. Li et al. (2017) previously reported that the hydrological observed precipitation is consistent with the meteorological observed precipitation.

Artificially observed soil moisture data from 2008 to 2010 were downloaded from the NMIC. The data were collected and measured artificially on the 8th, 18th, and 28th of every month, including 11 layers (0–5, 0–10/5–10, 10–20, 20–30, 30–40, 40–50, 50–60, 60–70, 70–80, 80–90, and 90–100 cm). The CLDAS-Prcp and CLDAS-V2.0 precipitation datasets were used to drive the Noah-MP model. The soil moisture depths simulated by Noah-MP were 0–10, 10–40, 40–100, and 100–200 cm, respectively. Only the 0–10-, 10–40-, and 40–100-cm soil moisture data were evaluated in this work to match the 242 soil moisture observation data after quality control.

The snow depth was defined as the vertical depth from the snow surface to the ground surface. The snow depth data are observations from > 2400 national meteorological stations archived by CIMISS.

Figure 3 shows the seasonal precipitation in the CMORPH, MERRA2, GLDAS-V2.1, and CLDAS-Prcp datasets. All four datasets show that precipitation in China decreases from southeast to northwest. The spatial distribution of precipitation agrees well among the four datasets in spring, summer, and autumn. However, precipitation in the CMORPH dataset (Fig. 3d) is significantly less than that in the other datasets in winter (Figs. 3h, l, p) in Northeast China, Xinjiang, and on the Tibetan Plateau. The CMORPH, GLDAS-V2.1, and CLDAS-Prcp (Figs. 3m–p) datasets show more detail in the distribution of precipitation than the MERRA2 dataset.

Figure 3.  Spatial distributions of seasonal precipitation (mm) for (a–d) CMORPH (1998–2018), (e–h) MERRA2 (1998–2018), (i–l) GLDAS-V2.1 (2000–2018), and (m–p) CLDAS-Prcp (1998–2018).

A total of 2380 values from the CMA precipitation observational dataset were used to evaluate the CLDAS-Prcp dataset. The data were used to calculate the bias, RMSE, and correlation coefficient, and these values were compared with the CMORPH, MERRA2, and GLDAS-V2.1 datasets. The results are shown as box plots that include the maximum, minimum, average, and median values and the outliers. Because the CLDAS-Prcp dataset used these site observations, it is referred to as the dependent evaluation.

Figure 4 shows box plots of bias, RMSE, and correlation coefficient for the CMORPH, MERRA2, GLDAS-V2.1, and CLDAS-Prcp datasets in spring from 2000 to 2018. The bias in the four precipitation datasets is small in spring. The CLDAS-Prcp dataset performs better than the other datasets and GLDAS-V2.1 shows the largest bias. The RMSE of the CLDAS-Prcp dataset varies from 0 to 7 mm day⁻¹ and the average is 2 mm day⁻¹, which is smaller than the RMSEs of the MERRA2 (0.2–11.5 mm day⁻¹; average 3.5 mm day⁻¹), CMORPH (1–13 mm day⁻¹; average 5.2 mm day⁻¹), and GLDAS-V2.1 (0.2–16 mm day⁻¹; average 5.7 mm day⁻¹) datasets. The box plots of the correlation coefficients (Fig. 4c) show that the CLDAS-Prcp dataset has the best correlation with a correlation coefficient > 0.8. The correlation coefficient of the CMORPH dataset is lower than that of the MERRA2 dataset, but better than that of the GLDAS-V2.1 dataset.

Figure 4.  Box plots of (a) bias, (b) RMSE, and (c) correlation coefficient for the CMORPH, MERRA2, GLDAS-V2.1, and CLDAS-Prcp datasets in spring from 2000 to 2018.

Figure 5 shows the box plots of the bias, RMSE, and correlation coefficient for the CMORPH, MERRA2, GLDAS-V2.1, and CLDAS-Prcp datasets in summer from 2000 to 2018. Figure 5a shows that the average bias is close to 0 mm day⁻¹ for the four precipitation datasets. The CLDAS-Prcp dataset performs better than the other datasets and GLDAS-V2.1 shows the largest bias. The RMSE of the CLDAS-Prcp dataset is from 1 to 8 mm day⁻¹ and the average is 5 mm day⁻¹, which is smaller than that of the MERRA2 (2.5–16 mm day⁻¹; average 9 mm day⁻¹), CMORPH (3.2–18 mm day⁻¹; average 10.5 mm day⁻¹), and GLDAS-V2.1 (3–18.5 mm day⁻¹; average 11 mm day⁻¹) datasets. The box plots of the correlation coefficients (Fig. 5c) show that the CLDAS-Prcp dataset has the best correlation with a correlation coefficient > 0.8. The correlation coefficient of the CMORPH dataset (average 0.6) is lower than that of the MERRA2 dataset (average 0.65), but better than that of the GLDAS-V2.1 dataset (average 0.45).

Figure 5.  As in Fig. 4, but for summer.

Figure 6 shows the box plots of the bias, RMSE, and correlation coefficient for the CMORPH, MERRA2, GLDAS-V2.1, and CLDAS-Prcp datasets in autumn from 2000 to 2018. Figure 6a shows that the average bias is close to 0 mm day⁻¹ for the four precipitation datasets. The CLDAS-Prcp dataset is better than the other datasets and GLDAS-V2.1 shows the largest bias. The RMSE of the CLDAS-Prcp is from 0 to 6 mm day⁻¹ and the average is 1.8 mm day⁻¹, which is smaller than that of the MERRA2 (0–11 mm day⁻¹; average 3 mm day⁻¹), CMORPH (0.5–13.5 mm day⁻¹; average 4.5 mm day⁻¹), and GLDAS-V2.1 (0–15 mm day⁻¹; average 4.7 mm day⁻¹) datasets. The box plots of the correlation coefficients (Fig. 6c) show that the CLDAS-Prcp dataset has the best correlation with a correlation coefficient > 0.8. The correlation coefficient of the CMORPH dataset (average 0.55) is lower than that of the MERRA2 dataset (average 0.7), but better than that of the GLDAS-V2.1 dataset (average 0.45).

Figure 6.  As in Fig. 4, but for autumn.

Figure 7 shows the box plots of bias, RMSE, and correlation coefficient for the CMORPH, MERRA2, GLDAS-V2.1, and CLDAS-Prcp datasets in winter from 2000 to 2018. Figure 7a shows that the CLDAS-Prcp dataset has the lowest bias, followed by the MERRA2 dataset. The average precipitation bias in the GLDAS-V2.1 dataset is better than that in the CMORPH dataset in winter. Figure 7b shows that the RMSE value of the CLDAS-Prcp dataset is from 0 to 1 mm day⁻¹ with an average of 0.5 mm day⁻¹, which is better than that of the MERRA2 (0–3.7 mm day⁻¹; average 0.8 mm day⁻¹), GLDAS-V2.1 (0–7 mm day⁻¹; average 1.8 mm day⁻¹), and CMORPH (0.5–8 mm day⁻¹; average 3.3 mm day⁻¹) datasets. The CLDAS-Prcp dataset has a higher correlation than the MERRA2 dataset, which, in turn, has a higher correlation than the GLDAS-V2.1 dataset. The CMORPH dataset has the lowest correlation coefficient because satellites have a lower ability to retrieve solid precipitation.

Figure 7.  As in Fig. 4, but for winter.

Because the CMA observational data for precipitation have been merged in the CLDAS-Prcp dataset, we selected the precipitation data from the MWR to evaluate the CLDAS-Prcp dataset and compared it with the GPM precipitation as an independent evaluation.

Figure 8 shows the time series of the bias, RMSE, and correlation coefficient for the GPM and CLDAS-Prcp datasets from 1 April to 31 October 2016–2018. Figure 8a shows that the bias of the CLDAS-Prcp dataset is better than that of the GPM dataset, which has a bias of −2.5 to 7.5 mm day⁻¹. The RMSE of the CLDAS-Prcp dataset is from 0 to 10 mm day⁻¹ with an average of 5.17 mm day⁻¹, whereas the RMSE of the GPM dataset is between 0 and 15 mm day⁻¹ with an average of 8.91 mm day⁻¹. The time series of the correlation coefficients (Fig. 8c) shows that the CLDAS-Prcp has the best correlation coefficient with the averege of 0.8, whereas the correlation coefficient of GPM is mostly about 0.6.

Figure 8.  Error time series of (a) bias, (b) RMSE, and (c) correlation of the GPM and CLDAS-Prcp datasets for 214 MWR precipitation sites from 1 April 2016 to 31 October 2018.

Because the CLDAS-V2.0 and CLDAS-Prcp datasets both include the CMA site data, we selected the MWR site data to evaluate these two datasets. However, as a result of the lack of MWR site data in winter, we can only use the CMA site data for the dependent tests. These two datasets are used to simulate the soil moisture content and snow depth in Noah-MP. The simulations were compared with in situ observations to indirectly evaluate the CLDAS-Prcp and CLDAS-V2.0 datasets.

Figure 9 shows the time series of the bias, RMSE, and correlation coefficient for the CLDAS-V2.0 and CLDAS-Prcp datasets from 1 April to 31 October 2016. Figure 9a shows that the bias of the CLDAS-Prcp dataset has mostly positive deviations, whereas the CLDAS-V2.0 dataset has mostly negative deviations. The RMSE (Fig. 9b) and correlation coefficients (Fig. 9c) of the CLDAS-Prcp dataset are similar to those of the CLDAS-V2.0 dataset. The spatial distribution of the RMSE (Fig. 10) shows that the CLDAS-Prcp dataset is also similar to the CLDAS-V2.0 dataset. In summary, the CLDAS-V2.0 and CLDAS-Prcp datasets are similar in summer.

Figure 9.  Error time series of (a) bias, (b) RMSE, and (c) correlation of the CLDAS-V2.0 and CLDAS-Prcp datasets for the MWR precipitation sites from 1 April to 31 October 2016.

Figure 10.  Spatial distributions of the RMSE of the (a) CLDAS-V2.0 and (b) CLDAS-Prcp datasets for 214 MWR precipitation sites from 1 April to 31 October 2016.

We calculated the RMSE against 2380 CMA daily stations to compare the CLDAS-V2.0 with CLDAS-Prcp dataset in winter. Figures 11a and 11b show the spatial distribution of the RMSE for the CLDAS-V2.0 and CLDAS-Prcp datasets in winter from 2013 to 2016. The RMSE values of the CLDAS-Prcp and CLDAS-V2.0 datasets decrease from southeast to northwest. The RMSE of the CLDAS-Prcp dataset is lower than that of the CLDAS-V2.0 dataset, except in Southeast and Southwest China. Figure 11c shows the statistical histogram of the RMSE for the CLDAS-V2.0 and CLDAS-Prcp datasets over different subregions in winter from 2013 to 2016. The CLDAS-Prcp dataset performs better than the CLDAS-V2.0 dataset in Northeast China, North China, Inner Mongolia, Xinjiang, and the Tibetan Plateau.

Figure 11.  Spatial distributions of the RMSE for the (a) CLDAS-V2.0 and (b) CLDAS-Prcp datasets, and (c) the statistical histogram of the RMSE in winter from 2013 to 2016.

Figure 12 shows the seasonal spatial distributions of the soil moisture content at 10-cm depth driven by the CLDAS-V2.0 and CLDAS-Prcp datasets. The soil moisture content driven by the CLDAS-V2.0 (CLDAS-V2.0-SM) and CLDAS-Prcp (CLDAS-Prcp-SM) datasets shows the spatial distribution of the soil moisture content in China. The CLDAS-V2.0-SM and CLDAS-Prcp-SM datasets are similar in most regions of China in spring, although the CLDAS-V2.0-SM dataset has lower values in Northeast China and northern Xinjiang. The CLDAS-Prcp-SM and CLDAS-V2.0-SM datasets in autumn and winter are similar to those in spring.

Figure 12.  Seasonal spatial distributions of the 0–10-cm soil moisture content simulated by the (a–d) CLDAS-V2.0 and (e–h) CLDAS-Prcp datasets from 2008 to 2010.

The RMSE and correlation coefficients over China were calculated by using observational data to quantitatively examine the performance of the CLDAS-V2.0-SM and CLDAS-Prcp-SM datasets. Figure 13 shows the time series of the RMSE and correlation coefficient of the 0–10-, 10–40-, and 40–100-cm soil moisture content in the CLDAS-V2.0-SM and CLDAS-Prcp-SM datasets from 2008 to 2010. The RMSE (Figs. 13a, c, e) shows that the CLDAS-Prcp-SM dataset performs better than the CLDAS-V2.0-SM dataset, especially for the 0–10-cm soil moisture content from January to March each year. The average RMSEs of different layers are 0.069, 0.067, and 0.072 mm³ mm⁻³ for the CLDAS-Prcp-SM dataset, whereas they are 0.073, 0.069, and 0.081 mm³ mm⁻³ for the CLDAS-V2.0-SM dataset. The correlation coefficients of the CLDAS-V2.0-SM and CLDAS-Prcp-SM datasets are similar, although the CLDAS-Prcp-SM dataset performs slightly better than the CLDAS-V2.0-SM dataset.

Figure 13.  Error time series of the (a, b) 0–10-, (c, d) 10–40-, and (e, f) 40–100-cm soil moisture driven by the CLDAS-V2.0 and CLDAS-Prcp datasets from 2008 to 2010.

Snow depth simulations of Noah-MP driven by the two datasets were compared with the observational data from November 2013 to April 2014 to investigate the impacts of the CLDAS-V2.0 and CLDAS-Prcp precipitation on the simulation of the land surface model.

Figure 14 shows the time series of the simulated snow depth driven by the CLDAS-V2.0 and CLDAS-Prcp precipitation. The snow depth from the CLDAS-Prcp precipitation forcing is close to that of the observational data from mid-November 2013 to late March 2014. The snow depth simulated by Noah-MP driven by the CLDAS-V2.0 (CLDAS-V2.0-snow) dataset is poorer than that driven by the CLDAS-Prcp (CLDAS-Prcp-snow) dataset, probably because the CLDAS-Prcp precipitation includes information about solid precipitation.

Figure 14.  Time series of the observational, CLDAS-V2.0, and CLDAS-Prcp snow depths.

Table 2 lists the evaluation results of the simulated snow depth averaged over Northeast China, Xinjiang, and the Tibetan Plateau from November 2013 to April 2014. Table 2 clearly shows that the CLDAS-Prcp-snow simulation is better than the CLDAS-V2.0-snow simulation over Xinjiang and Northeast China, whereas the relative bias of the CLDAS-Prcp-snow (−54.99%) dataset is larger than that of the CLDAS-V2.0-snow (−36.98%) dataset over the Tibetan Plateau. The CLDAS-Prcp-snow dataset performs better than the CLDAS-V2.0-snow dataset from the perspective of the correlation coefficient and relative RMSE. Overall, the CLDAS-Prcp-snow dataset performs better than the CLDAS-V2.0-snow dataset because CLDAS-Prcp includes information about solid precipitation.

There are no hourly observations of solid precipitation over the icy regions of China except for weighing-type precipitation sensors and satellite microwave retrieval data. We blended CMORPH and MERRA2 precipitation datasets with observational temperature and precipitation data on various temporal scales to produce the CLDAS-Prcp precipitation dataset using the STMAS and temporal downscaling methods. The CLDAS-Prcp dataset was then evaluated from the perspective of the total seasonal precipitation, dependent evaluation, independent evaluation, and land surface model simulation. Our major conclusions are as follows.

(1) The temporal downscaling method was combined with a temperature threshold to produce a new hourly precipitation dataset based on MERRA2 and daily and hourly precipitation data. This new dataset compensates for missing observations of the hourly precipitation over icy areas where only non-weighing precipitation sensors that cannot measure solid precipitation are deployed. Hourly solid precipitation observations will be available in the future as more weighing-type precipitation sensors are deployed. We used the 2.5°C threshold of Jordan (1991) to distinguish between rainfall and solid precipitation, whereas the Multi-Source Weighted-Ensemble Precipitation dataset used a 3°C threshold (Beck et al., 2017). We plan to determine the threshold in different regions of China by comparing weighing-type precipitation sensors with the traditional rain gauges.

(2) The multiyear average precipitation shows that the CLDAS-Prcp dataset can reasonably reflect the spatial distribution of precipitation in China. However, because the CMOPRH precipitation in lakes on the Qinghai–Tibetan Plateau is greater than that in the surrounding areas, this also occurs in the CLDAS-Prcp dataset. The specific reason for this requires further analysis. The dependent evaluation using observational data from the CMA indicates that the CLDAS-Prcp dataset performs better than the CMORPH, MERRA2, and GLDAS-V2.1 datasets. The CLDAS-Prcp dataset also performs better than the GPM precipitation based on an independent evaluation using MWR observational data. This is because the CLDAS-Prcp dataset merges more observational precipitation data.

(3) Compared with the CLDAS-V2.0 precipitation, the CLDAS-Prcp dataset performs better in the dependent evaluation in winter using observational data from the CMA, whereas the CLDAS-V2.0 and CLDAS-Prcp datasets are similar in an independent evaluation by the MWR observational data in summer. From the perspective of land surface model simulation, the soil moisture content simulated by the CLDAS-V2.0 dataset is lower than that simulated by the CLDAS-Prcp dataset in Northeast China and northern Xinjiang in all seasons except summer. The soil moisture content simulated by the CLDAS-Prcp dataset performs better than the CLDAS-V2.0 dataset, especially at 0–10-cm depth from January to March. This is mainly because the solid precipitation in the CLDAS-Prcp dataset leads to an improved soil moisture simulation in the following period. The snow depth simulated by the CLDAS-Prcp dataset is much better than that simulated by the CLDAS-V2.0 dataset. This is mainly because the CLDAS-Prcp dataset includes information about solid precipitation. The CLDAS-Prcp dataset can therefore be used to drive land surface models.

In summary, CLDAS-Prcp well reflects the spatial distribution of precipitation in China. As an important component of CLDAS, the CLDAS-Prcp dataset is expected to be widely used in studies of land surface hydrological processes and the climate in China. Radar-observed precipitation, more satellite precipitation, reanalysis precipitation, wind speed, and wind direction data will be blended into this dataset. Possible problems in this dataset will be found and fixed during its application to further improve its quality.